An AI agent that can call tools — send emails, update records, move money, trigger workflows — is no longer just answering questions. It is acting. That shift is where most of the real risk in agentic AI systems lives: not in what the model says, but in what it is allowed to do, and under what conditions.
AI Agent Security is a focused review of an autonomous agent's permission model, approval gates, and blast radius — distinct from the general AI Security Kenya review and from the workflow-level Secure Automation Review.
Who this is for
- Teams deploying AI agents that call external APIs, write to databases, or trigger business processes.
- Products where an agent acts on behalf of a customer or employee without a human confirming every step.
- Any system where an agent's tool access, if abused or manipulated, could cause real financial or operational harm.
What gets reviewed
| Review Area | What Is Examined |
|---|---|
| Tool-Call Permission Boundaries | Whether each tool the agent can call is scoped to the minimum access it needs. |
| Human-Approval Gates | Whether consequential actions (payments, deletions, external communication) require explicit confirmation. |
| Trajectory Auditing | Whether the agent's decision path and tool calls are logged in a way a human can review after the fact. |
| Sandboxing & Blast Radius | What the worst-case outcome is if the agent is manipulated into taking a harmful action, and how contained it stays. |
| Input Validation | Whether tool-call arguments are validated against a schema before execution, not trusted as generated. |
The boundary: permission before testing
Complementary engineering work
This review pairs naturally with the agent-engineering practices documented by HarLyn Digital Partners' AI Systems Lead, Harrison Ndeke, including his field notes on validating AI agent tool calls against JSON schema and human-approval boundaries, and his AI Agent Developer Kenya service.
What you receive
- Findings register: permission and approval-gate issues with severity and reproduction evidence.
- Permission map: what each tool the agent can call is actually capable of, verified against what it needs.
- Remediation guidance: specific fixes for scoping, approval gates, and audit logging.
Evidence and method
Ongoing detection and response capability relevant to agent-permission auditing is demonstrated in HarLyn Sentinel.
Frequently asked questions
What's a human-approval gate, concretely?
A checkpoint that pauses an agent before it takes an irreversible, financially material, or externally-visible action — requiring explicit human confirmation before the agent proceeds, rather than acting autonomously.
How do you test an agent's tool permissions?
By mapping what each tool the agent can call is actually capable of against what the agent's task requires — verified, not assumed from the configuration — and attempting to trigger out-of-scope tool calls under controlled conditions.
Can an AI agent's actions be traced back after the fact?
Only if audit logging captures the agent's reasoning trail and tool calls at the time of execution — this review checks whether that trail exists and is actually queryable, not just whether logging is enabled in principle.
Does this apply to agents built on any framework, or only specific ones?
The review method is framework-agnostic — n8n-orchestrated agents, LangChain, custom tool-calling loops, and similar architectures are all in scope.
How engagements start: the 48-Hour Secure Digital Workflow Assessment
Most clients begin with the 48-Hour Secure Digital Workflow Assessment, delivered jointly with HarLyn Digital Partners, which identifies whether a dedicated agent security review is warranted.
Engagements begin with a fixed-scope scoping call. Clear decisions before code, and no obligation to proceed to follow-on build work.
Request a Scoping Call →Related: RAG Security · AI Security Kenya · Secure Automation Review · All services